On November 9 Anthropic published a research post in which Claude Fable 5.2 took three electronic devices from design to working prototype: a USB speakerphone, a home voice-assistant hub and a handheld virtual pet. The partner was Diode Computers, a circuit board design and manufacturing company Anthropic has worked with since 2025 and has a commercial relationship with.
The sentence that matters: Diode manufactured all three boards straight from Claude’s design files, with no rounds of adjustments. The team says it was surprised, because those rounds are normally part of the job.
What Claude delivered per device
Each project started from a short prompt. For the speakerphone it asked for “a puck that plugs into a computer over USB-C and includes microphones and a speaker, supporting Bluetooth with a PCB antenna.” Out of that came the specification, the schematics, component placement, the copper routing, a 3D-printable enclosure, the firmware and a parts list. Diode’s engineers assembled the boards and ran the firmware, under Claude’s direction.
Geometry was the jump
Anthropic puts the biggest difference from earlier models in geometric reasoning. Claude routed the important connections first — high-speed signals, high currents, noise-sensitive lines — and kept differential pairs matched in length and spacing, including the zigzag tuning patterns that even out arrival times.
On Diode’s internal PCB Gym benchmark, which asks models to place components and reroute connections, Fable 5.2 scores 69% against 25% for Fable 5. On the public BenchCAD 2.1 it reaches 83.5%, nearly double Fable 5.1.
Two places where Claude left its tools behind
On the voice-assistant hub, long traces ran off a 100-pin connector. KiCad’s router draws one trace between two points Claude picks, which is too granular at that volume. So Claude wrote its own router: an algorithm that divides the board into a grid and searches it for a clear path. Within about half an hour each long trace took roughly a second. When an audio clock trace sat too close to the Ethernet pairs, Claude adjusted its own router to leave more room around sensitive lines.
The second place is the enclosure. Claude calculated that a 25 mm microphone port would resonate near 3.4 kHz, right in the voice band, and designed a wedge-shaped case that shortens the port to about 10 mm and pushes the resonance to roughly 8.9 kHz. For the speakerphone it found in its own 3D model that the board cannot drop straight into the case because the mute switch hits the back wall, so it added a step to the assembly instructions: tilt the board about 25 degrees.
What Anthropic rules out explicitly
All three devices are prototypes. They have not been through electrical validation, antenna tuning, thermal testing or certification, and they were not designed for mass production. Claude was not connected to test equipment or factory machines; people built and tested them. Anthropic also points out that OpenAI recently wrote about using its own models to design a handheld device.
The checking step is the new part
The manufacturing checks are the part I read twice. Claude ran the DFM and sourcing checks itself and sorted the results: six slivers of solder mask flagged as too thin to manufacture, which it fixed by widening the openings without moving any parts. A flagged connection between two grounds it recognized as deliberate and ignored. Anyone who has sent a board to a factory knows that this sorting is the actual work, and that a machine getting it wrong costs an expensive turnaround.